Interpretable Deep Learning for Risk Representation: A Risk Model for Construction Safety

Abstract

Traditional accident models in construction safety either rely heavily on expert judgment or employ black-box machine learning (ML) approaches that lack interpretability. This study proposes an interpretable deep learning framework for risk representation. A unified holistic risk variable is introduced to integrate heterogeneous contributing factors, including human, environmental, organizational, project, hazard, and violation elements, into a normalized [0,1] scale. Based on 120 fatal accident investigation reports, structured features are extracted and mapped into this risk abstraction. A two-stage modeling strategy is employed: an artificial neural network (ANN) captures nonlinear relationships among risk factors, and a long short-term memory (LSTM) network models the temporal evolution of the holistic risk state. To enhance transparency, a global linear regression model and a local interpretable model-agnostic explanation (LIME) model provide complementary global and local explanations of modeled risk dynamics. Results demonstrate that the framework captures nonlinear and sequential patterns while maintaining interpretability, offering a methodological foundation for dynamic and transparent safety analytics.